Comparing risk adjustment estimation methods under data availability constraints

The Italian National Healthcare Service relies on per capita allocation for healthcare funds, despite having a highly detailed and wide range of data to potentially build a complex risk-adjustment formula. However, heterogeneity in data availability limits the development of a national model. This paper implements and evaluates machine learning (ML) and standard risk-adjustment models on different data scenarios that a Region or Country may face, to optimize information with the most predictive model. We show that ML achieves a small but generally statistically insignificant improvement of adjusted R2 and mean squared error with fine data granularity compared to linear regression, while in coarse granularity and poor range of variables scenario no differences were observed. The advantage of ML algorithms is greater in the coarse granularity and fair/rich range of variables set and limited with fine granularity scenarios. The inclusion of detailed morbidity- and pharmacy- based adjustors generally increases fit, although the trade-off of creating adverse economic incentives must be considered.